// project
A Hinglish social-deduction bluffing game of power, satire & second chances. Gaddi ke liye kuch bhi karega.
The brief
Gaddi is a Hinglish social-deduction bluffing game set in a satirical India corporate-political underworld where six archetypes scheme for an empty chair, the Gaddi, and everyone is lying about what they hold. The Neta makes promises he'll forget tomorrow, the Bhai owns silence, the Babu approves nothing, the Jugaadu knows a shortcut, the Vakil has read every exception. Satire targets the archetype, never the person. Under the deadpan Hinglish voice ("सब मिले हुए हैं") sits a serious engineering exercise: one deterministic Kotlin engine that runs identically on Android, iOS, desktop and the web, and powers the AI, the UI and a server-authoritative backend from the same code.
What shipped
no public push in the last 20 events
// case study
The problem
A hidden-information bluffing game handed every player the same expert board on turn one, suspicion odds and a live teleprinter log included, which is exactly the kind of interface that loses a first-timer before their first bluff.
The decision
Split the board into three density layers a player graduates through by playing, not a settings toggle: FOCUS shows only whose turn it is, one plain-language line, and your legal moves; GUIDED adds coaching; ANALYST is the full instrument panel.
The result
All three layers, the AI, and a future server read the same deterministic (GameState, Intent) → GameState engine: the density decision is a UI skin, not a fork in the rules.
4 platforms · one engine
Watch the two-minute tour ↓
Analyst view: Gaddi draws on 6 stack entries and links to 4 outside sources, and shares 2 shared foundations with 3 other builds.
// guided tour
A recorded walkthrough of the app with sample content. Tap the speaker for the voiceover, or read the captions. In-app figures describe the recorded demo, not live production activity.
// shipping
Every line below is checkable. Add the repository in an F-Droid client and the app installs, signed with the same key its listing pins.
// pipeline
Read live from GitHub Actions and the F-Droid index. Not a badge: the actual runs, what shipped, and the certificate you can check the download against.
reading the pipeline
4
platforms · one engine
10
AI bot personas
6
roles · 4 story arcs
7
toggle rule variants
// multiplatform
Recorded app screens and, where available, the running build. Sample content inside a screen is separate from the source-linked project metrics.
Rendered at phone dimensions from the shared Compose UI.
// design notes
The whole game is a pure function: (GameState, Intent) → GameState, with the RNG seed living inside the state. The same module drives single-player, the bots and a future server. Any match can be replayed byte-for-byte from its seed and intent log.
The board reveals itself in three density layers so a first-timer isn't handed an expert's dashboard. FOCUS shows only whose turn it is, one plain-language line of what just happened, your hand and your legal moves; GUIDED adds gentle coaching; ANALYST is the full instrument panel (suspicion pips, odds, teleprinter log). Players graduate FOCUS → GUIDED → ANALYST by playing. Paired with a tap-to-continue beat gate so the round never resolves faster than you can read it, and a tutorial-first onboarding funnel that teaches one mechanic at a time.
A diegetic court-scribe turns raw engine events into one calm in-character line, grounded strictly on the redacted PlayerView so it narrates the beat without ever leaking a hidden card or inventing the board. It renders the deterministic templated line instantly and upgrades in place if an LLM is available (on-device Gemini Nano / Apple FoundationModels / BYOK cloud), never enters the intent log, and leaves byte-for-byte replay untouched.
Bots use Information Set Monte Carlo Tree Search (1.5k-16k iterations depending on difficulty tier) with an optional cloud-LLM upgrade (Anthropic / OpenAI / Gemini). Built on kmp-toolkit's own generic bots-policy shell (Policy/GameRules/Ismcts/SearchBudget, zero deps), not a bespoke search implementation. Ten personas each have a personality profile driving targeting and bluff frequency. The DARBAR layer lets bots form alliances, hold grudges and trade Hinglish table-talk across four story arcs. Social manipulation that never breaks engine determinism.
A hidden-information game needs strict secrecy: redact(state, viewer) → PlayerView guarantees a client only ever sees what its player should. Two independent narrative RNG streams keep flavour separate from game logic.
A bespoke lamplit visual language: teak/brass/cream palette, Rozha One display type, Canvas-drawn intaglio role glyphs and stamped-instrument motifs. The launch overhaul pushed it to an AAA bar: every screen dissolved from bordered boxes into one continuous lit table (depth via shadow, never outline), a shared component vocabulary, and an AGSL/Skia runtime-shader material layer (film grain + warm bloom on the felt) with a graceful no-shader fallback. All behind a full Fastlane + CI pipeline with headless screenshot rendering.
New Game (1v1-1v9, Easy→Grandmaster), a KISSA story campaign, GAUNTLET (Tarakki ki Seedhi, a 5-rung ladder ending in 6-player Grandmaster), TAMASHA (spectate ten AI personas scheme and betray), Team Khel (faction play with un-targetable allies), a Tutorial you can't leave until you catch a bluff, local pass-and-play with a handoff screen guard, and online + LAN multiplayer.
Four narrative arcs run at once, fuelled or suppressed by your chat suggestions: GATHBANDHAN (a quiet coalition, watch who breaks first), AFWAAH (a rumour the table acts on even when false), STING (a leaked claim that forces a read), and BADLA (a vendetta that outlives the round). They run on a separate deterministic narrative RNG that never touches card state and resumes byte-for-byte.
All six roles use the Okabe-Ito colourblind-safe palette plus a unique engraved bezel pattern (ring, hatch, dots, weave, double-rule, ticks) so identity reads without colour. Reduced-motion mode swaps every beat for a bespoke static end-frame (GHOTALA = held stamp, SUPARI = tipped chair). Accessibility never flattens the narrative.
An AiProvider interface abstracts Anthropic, OpenAI, Gemini, on-device Gemini Nano (Android) and Apple FoundationModels (iOS 26); ISMCTS is the always-available offline fallback. Bring-your-own-key, stored in each platform's encrypted storage.
Online and LAN play (private room codes, quick-match, Bonjour/mDNS discovery) run on a Ktor/Netty server that holds all state; clients receive only their redacted PlayerView, so another player's face-down roles can't appear on the wire by construction.
Seven additive rule variants (Bail Pe Bahar, Bali Khel, Hawala, Adhyadesh, Khazana Raj, Mehengai, Tangi) combine freely and default off. The engine is byte-for-byte unchanged when they're disabled, expanding the surface without touching core logic.
// cast
The Politician: Tax +3 (GHOTALA); blocks Foreign Aid
The Don: Assassinate −3 (SUPARI); unblockable except by the Vakil
The Bureaucrat: Steal 2 (VASOOLI); blocks Steal
The Fixer: Exchange cards (SETTING); blocks Steal
The Lawyer: no action; blocks Assassinate only (power through procedure)
The Journalist: Investigate a card (JAANCH); unblockable
// architecture
// under the hood
// directions
Pick a treatment to see it whole. Where one ships a dark counterpart, drag the divider to compare the two faces of it — same layout, same content, only the light changes.
// gallery
Swipe or use the arrows · tap a screen to enlarge
Follow the connections
This project uses kmp-build-logic and kmp-toolkit. Follow the same foundations into another build.